A disease group data structure analysis method and system based on big data

Abstract

The application provides a disease group data structure analysis method and system based on big data, which obtains disease group data of a user, processes the disease group data in a structured and unstructured manner, splices the data through a cross-modal attention mechanism to form multi-modal features, determines disease course stratification according to clinical guideline stage division and data-driven risk inflection points, and constructs a disease course stratification network, wherein each layer of the network comprises nodes specific to the corresponding stage; the adjacency matrix of the disease course stratification network, the multi-modal features and timestamp features are input into a multi-modal spatio-temporal graph neural network model to output multi-modal correlation scores of each node and core dependency paths of each layer of the network; the nodes of the disease course stratification network are constrained according to clinical knowledge, and target paths in the core dependency paths are screened out; the target paths are subjected to counterfactual simulation verification, and a final path that passes the verification is output, and the final path can directly support personalized diagnosis and treatment decisions.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of disease group data structure analysis, and particularly relates to a disease group data structure analysis method and system based on big data. BACKGROUND

[0002] In clinical diagnosis and treatment and medical management, accurately mining the internal dependence relationship among diagnosis, treatment and outcome is a core requirement for optimizing clinical path, improving treatment effect and reasonably allocating medical resources.

[0003] In the current medical data analysis field, the correlation analysis of disease groups mainly depends on static diagnosis grouping systems (such as DRG / DIP) or simple event co-occurrence statistics. Such methods have significant technical limitations:

[0004] Firstly, the time evolution characteristics of medical events are ignored, and the dynamic correlation (such as the delayed influence of specific treatment in the acute phase on the follow-up period outcome) of diagnosis, treatment and outcome in different stages (such as acute phase and recovery period) cannot be captured.

[0005] Secondly, the hierarchical nested relationship (such as time, method, complications and other sub-dimensions contained in the treatment plan) of diagnosis and treatment events is not considered, resulting in that the correlation analysis only stays at the macro level and it is difficult to mine the key influence of sub-dimensions on the outcome.

[0006] Thirdly, only statistical correlation between events can be identified, and the interference of confounding factors such as age and underlying disease cannot be excluded, it is difficult to distinguish between “correlation” and “causality”, and misleading conclusions (such as mistaking the treatment with co-occurrence but no actual causality with the outcome as a key intervention method) are easy to produce.

[0007] These defects make the existing technology unable to effectively mine the hidden dependence relationship of “diagnosis-treatment-outcome” in disease groups, and it is difficult to meet the actual needs of clinical precise diagnosis and treatment decision, prognosis evaluation and risk warning. Therefore, a disease group correlation analysis technology integrating time dynamics, hierarchical structure and causal reasoning is urgently needed. SUMMARY

[0008] Therefore, in the embodiments of the present application, a disease group data structure analysis method and system based on big data are provided, which aims to convert disease group data from a classification structure to a network structure, and focuses on analyzing the correlation strength and causal relationship between different medical elements, rather than a single dimension statistics, to provide direct basis for clinical path optimization.

[0009] The first aspect of the embodiment of the present application provides a disease group data structure analysis method based on big data, which comprises:

[0010] Obtaining disease group data of a user, performing structured feature processing and unstructured feature processing on the disease group data, and splicing results of the structured feature processing and the unstructured feature processing through a cross-modal attention mechanism to form multi-modal features;

[0011] According to the clinical guideline stage division and the data-driven risk inflection point, the disease course stratification is determined, and a disease course stratification network is constructed, each layer network of the disease course stratification network comprising nodes exclusive to the corresponding stage, the nodes including diagnosis nodes, treatment nodes and outcome nodes;

[0012] The adjacency matrix of the disease course stratification network, the multi-modal features and the timestamp features are input into a multi-modal spatio-temporal graph neural network model, and the multi-modal correlation scores of each node and the core dependency paths of each layer network are output;

[0013] According to clinical knowledge, the nodes of the disease course stratification network are constrained, and the target path in the core dependency path is screened out;

[0014] The target path is verified by counterfactual simulation, and the final path that passes the verification is output.

[0015] Further, in the step of obtaining disease group data of a user, performing structured feature processing and unstructured feature processing on the disease group data, and splicing results of the structured feature processing and the unstructured feature processing through a cross-modal attention mechanism to form multi-modal features, in the process of structured feature processing, ICD-10 diagnosis codes, ATC medication codes and outcome indicators are converted into 64-dimensional feature vectors through encoding embedding and numerical normalization processing;

[0016] In the process of unstructured feature processing, a medical pre-training model is used to extract lesion features to generate a 256-dimensional image feature vector;

[0017] In the process of unstructured feature processing, core gene sites related to diseases are screened through gene pathway analysis and converted into 128-dimensional gene feature vectors;

[0018] In the process of unstructured feature processing, a time convolution network is used to extract rhythm features to generate a 64-dimensional physiological feature vector.

[0019] Further, the step of determining the disease course stratification according to the clinical guideline stage division and the data-driven risk inflection point comprises:

[0020] According to the clinical guidelines, the stage division is performed to obtain a preliminary stratification result;

[0021] The historical data is subjected to LSTM-based time series inflection point identification to correct the boundaries of the preliminary stratification result, and the disease course stratification is determined.

[0022] Further, different hierarchical nodes are connected by stage transition edges, wherein the edge weight of the stage transition edge is the contribution degree of the corresponding treatment pair to the stage transition, which is calculated by Cox regression.

[0023] Further, the core layer of the multi-modal spatio-temporal graph neural network model comprises time feature learning, spatial correlation learning and cross-modal fusion, in the time feature learning, 1D convolution is adopted to extract the time sequence dependence of node events, and a time sequence feature vector is output, in the spatial correlation learning, the spatial correlation of nodes in the same layer / cross layer is learned based on a graph attention network, and a multi-modal attention weight is introduced, and a spatial feature vector is output, in the cross-modal fusion, the time sequence feature vector and the spatial feature vector are combined through a gating fusion unit.

[0024] Further, in the step of performing constraint on the nodes of the disease course hierarchical network according to clinical knowledge, and screening the target path in the core dependency path, first, the causal relationship of the node pair in the disease course hierarchical network is marked, when the causal relationship is clear causality, positive constraint is performed, that is, the initial value of the corresponding edge weight is improved, and when the causal relationship is clear irrelevant, reverse constraint is performed, that is, the upper limit of the corresponding edge weight is set to filter pseudo-association.

[0025] Further, the step of performing counterfactual simulation verification on the target path and outputting a final path that passes verification comprises:

[0026] obtaining a clinical contraindication library, and verifying whether the virtual intervention scheme for the target path conforms to the clinical operation specification;

[0027] If the virtual intervention scheme for the target path conforms to the clinical operation specification, the influence of the intervention on the outcome is verified by calculating the intervention effect, specifically, for the non-continuous variable intervention, the absolute change value and the relative change percentage of the outcome after intervention are calculated, and it is judged whether the absolute change value and the relative change percentage are greater than the corresponding threshold value, and at the same time, for the continuous variable intervention, it is verified whether the intervention intensity and the outcome improvement are positively correlated / staircase correlated;

[0028] If the absolute change value and the relative change percentage are greater than the corresponding threshold value, and the intervention intensity and the outcome improvement are positively correlated / staircase correlated, the stability of the intervention effect under different conditions is verified through sensitivity analysis, specifically, whether the effect fluctuation amplitude of the intervention effect recalculated after adjustment of confounding factors is less than the preset fluctuation amplitude is verified, and whether Bootstrap sampling verification is up to standard is verified.

[0029] If the effect fluctuation amplitude of the recalculated intervention effect after verification of confounding factors is less than the preset fluctuation amplitude, and the Bootstrap sampling verification meets the standard, whether the causal logic of the intervention and the outcome is contradictory is verified according to the medical pathological mechanism and the pharmacological principle.

[0030] If the causal logic of the intervention and the outcome is not contradictory, the corresponding path is output.

[0031] The second aspect of the embodiment of the application provides a disease group data structure analysis system based on big data, which is used for realizing the disease group data structure analysis method based on big data in the first aspect, and the system comprises:

[0032] The feature processing module is configured to acquire disease group data of a user, perform structured feature processing and unstructured feature processing on the disease group data, and splice the results of the structured feature processing and the unstructured feature processing through a cross-modal attention mechanism to form multi-modal features.

[0033] The construction module is configured to determine disease course stratification according to clinical guideline stage division and data-driven risk inflection points, and construct a disease course stratification network, wherein each layer network of the disease course stratification network comprises nodes exclusive to corresponding stages, and the nodes include diagnosis nodes, treatment nodes and outcome nodes.

[0034] The input module is configured to input an adjacency matrix of the disease course stratification network, the multi-modal features and timestamp features into a multi-modal spatio-temporal graph neural network model, and output multi-modal correlation scores of each node and core dependency paths of each layer network.

[0035] The screening module is configured to constrain the nodes of the disease course stratification network according to clinical knowledge, and screen out target paths in the core dependency paths.

[0036] The verification module is configured to perform counterfactual simulation verification on the target paths, and output a final path that passes the verification.

[0037] The third aspect of the embodiment of the application provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the disease group data structure analysis method based on big data provided in the first aspect.

[0038] The fourth aspect of the embodiment of the application provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor realizes the disease group data structure analysis method based on big data provided in the first aspect when executing the program.

[0039] The embodiment of the present application provides a disease group data structure analysis method and system based on big data, through obtaining disease group data of a user, structured feature processing and unstructured feature processing are performed on the disease group data, and through a cross-modal attention mechanism, results of the structured feature processing and the unstructured feature processing are spliced to form multi-modal features; according to clinical guideline stage division and data-driven risk inflection points, disease course stratification is determined, and a disease course stratification network is constructed, each layer of the disease course stratification network comprises nodes exclusive to a corresponding stage, and the nodes include diagnosis nodes, treatment nodes and outcome nodes; an adjacency matrix of the disease course stratification network, the multi-modal features and timestamp features are input into a multi-modal spatio-temporal graph neural network model, and multi-modal correlation scores of each node and core dependent paths of each layer of the network are output; according to clinical knowledge, nodes of the disease course stratification network are constrained, and a target path in the core dependent path is screened out; the target path is subjected to counterfactual simulation verification, and a final path that passes verification is output. Specifically, the present application breaks through the limitation of a single-modal static network, fuses multi-dimensional medical data, constructs a disease course dynamic stratification, multi-modal feature fusion and clinically knowledge-constrained causal reasoning system, realizes the leap from data correlation to clinically interpretable causal dependence, and directly supports personalized diagnosis and treatment decisions. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 An implementation flowchart of the disease group data structure analysis method based on big data provided by the embodiment one of the present application is provided.

[0041] Figure 2 A structural block diagram of the disease group data structure analysis system based on big data provided by the embodiment two of the present application is provided.

[0042] Figure 3 A structural block diagram of an electronic device provided by the embodiment three of the present application is provided. DETAILED DESCRIPTION

[0043] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the related drawings. The drawings show several embodiments of the present application. However, the present application can be realized in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0044] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there can be a middle element. When an element is referred to as being "connected" to another element, it can be directly connected to the other element or there can be a middle element. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. The use herein of the terms "and / or" includes a set of one or more associated listed items.

[0046] Embodiment one

[0047] According to the embodiments of the present application, a disease group data structure analysis method based on big data is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0048] In this embodiment one, a disease group data structure analysis method based on big data is provided, which can be used in electronic devices such as computers. Please refer to Figure 1 , Figure 1 The implementation flowchart of the disease group data structure analysis method based on big data provided by the embodiment one of the present application is shown, which specifically includes steps S01 to S05.

[0049] In step S01, the disease group data of the user is obtained, and the disease group data is processed by structured feature processing and unstructured feature processing, and the results of the structured feature processing and the unstructured feature processing are spliced by a cross-modal attention mechanism to form multi-modal features.

[0050] Specifically, the ICD-10 diagnosis code, ATC medication code, and outcome indicator are converted into a 64-dimensional feature vector through coding embedding and numerical normalization processing, wherein the outcome indicator can be blood pressure, blood glucose, etc. For example, type 2 diabetes is mapped to a code vector, and fasting blood glucose 8.5 mmol / L is normalized to [0, 1];

[0051] In the process of unstructured feature processing, a medical pre-training model is used to extract lesion features to generate a 256-dimensional image feature vector. This process is for processing image data such as CT, MRI, etc. The medical pre-training model can be CheXNet or MONAI, and the lesion features include tumor size, edema degree, etc.

[0052] In the process of unstructured feature processing, the core gene sites related to the disease are screened through gene pathway analysis to convert into a 128-dimensional gene feature vector. This process is for processing gene data such as tumor driver genes and metabolic disease susceptibility genes. The gene pathway analysis is KEGG pathway enrichment.

[0053] In the process of unstructured feature processing, a time convolution network is used to extract rhythm features to generate a 64-dimensional physiological feature vector. The process is for processing physiological signals such as electrocardiogram, electroencephalogram, etc., wherein the rhythm features include heart rate variability, proportion of alpha waves in electroencephalogram, etc.

[0054] More specifically, in the process of forming a multi-modal feature by splicing the results of structured feature processing and unstructured feature processing through a cross-modal attention mechanism, the structured feature processing result is obtained, and the unstructured feature processing result is compressed to the same dimension. Combined with clinical priori (such as giving higher basic weight to image features in the tumor scenario, and focusing on structured blood glucose indicators in the diabetes scenario) and data-driven signals (calculating the similarity between modalities, the confidence of unstructured features such as image lesion confidence), the dynamic attention weight of each modality is obtained through Softmax normalization; Finally, the two types of features are weighted and summed according to the weight, and a cross-modal feature crossing item (such as the element interaction between structured pathological staging and image lesion features) can be added to enhance semantic association, and finally a cross-modal fusion feature is generated that takes into account clinical significance and data integrity.

[0055] It can be understood that by including images, genes, and physiological signals into the disease group association network, the node feature dimension is enriched, and the problem of single data dimension in traditional solutions is solved.

[0056] Step S02, according to the clinical guideline stage division and data-driven risk inflection point, determine the disease course stratification, and construct a disease course stratification network, each layer of the disease course stratification network contains nodes exclusive to the corresponding stage, the nodes include diagnosis nodes, treatment nodes and outcome nodes.

[0057] Specifically, according to the clinical guidelines, stage division is performed to obtain a preliminary stratification result. It can be understood that what is actually done is time stratification, i.e., dividing time layers according to the patient's disease course stage. For example, according to the guidelines for the prevention and treatment of type 2 diabetes, type 2 diabetes can be divided into insulin resistance stage, impaired glucose tolerance stage, clinical diabetes stage, and complication stage.

[0058] The historical data is subjected to LSTM-based time sequence inflection point identification, the boundary of the preliminary stratification result is corrected, and the disease course stratification is determined. It should be noted that in the process of LSTM-based time sequence inflection point identification, the historical data, i.e., medical time sequence data (such as daily blood glucose value, heart rate change, inflammation index dynamic, etc.) is first divided into input sequences according to a fixed time window, the time sequence characteristics and potential change patterns of the sequences are learned through the encoder of the LSTM network, and then the original sequences are reconstructed through the decoder, and the reconstruction error is calculated. When the reconstruction error at a certain time is significantly higher than the set threshold, and combined with the verification of clinical field knowledge, for example, the key events such as treatment scheme adjustment and complication occurrence corresponding to the time, it can be determined that the time is a time sequence inflection point. In addition, the attention mechanism can be added to the LSTM model to automatically focus on the time step data that contributes more to the trend change, further improving the accuracy of the inflection point identification, and finally providing data-driven time node basis for disease group dynamic stratification and disease progression stage division. For example, it is found that glycosylated hemoglobin ≥ 9.0% is the risk inflection point of clinical diabetes period → complication period, which is set as the stratification threshold.

[0059] It should be noted that different levels of nodes are connected by phase transition edges, for example, a cross-layer edge is established between clinical diabetes period → SGLT2 inhibitor treatment and complication period → kidney function stability, wherein the edge weight of the phase transition edge is the contribution degree of the corresponding treatment to the phase transition, which is calculated by Cox regression. In addition, each node is attached with a multi-modal feature label, for example, the diabetic nephropathy diagnosis node is attached with kidney ultrasound features and eGFR gene site features.

[0060] In step S03, the adjacency matrix of the disease course stratification network, the multi-modal features and the timestamp features are input into a multi-modal spatio-temporal graph neural network model, and the multi-modal association scores of each node and the core dependence paths of each layer network are output.

[0061] In the embodiments of the present application, the adjacency matrix (including inter-layer edges) is a block matrix structure with directional weights, the core of which is to divide the matrix block according to the disease course stage (such as “pre-hospital emergency period-in-hospital acute period-recovery period-follow-up period” of myocardial infarction, “insulin resistance period-clinical diabetes period-complication period” of diabetes), and at the same time, the intra-layer node association and inter-layer node association are described through a dedicated sub-matrix, realizing the structured expression of the dynamic dependence of “diagnosis-treatment-outcome”.

[0062] Specifically, the matrix is divided into several diagonal blocks and non-diagonal blocks according to the level: the diagonal blocks correspond to the same layer adjacent sub-matrix of each layer, and the elements in each sub-matrix represent the correlation weight between the nodes in the layer (such as acute phase, myocardial infarction diagnosis, PCI treatment, shock outcome), the weight value is calculated by integrating the co-occurrence frequency of nodes in the layer, the clinical effect quantity (such as the relative risk ratio RR of treatment and outcome) and the multi-modal feature correlation degree (such as the matching degree of imaging lesion features and treatment response), and the non-zero value indicates that the nodes are associated, and the larger the value, the stronger the association; the non-diagonal blocks correspond to the inter-layer adjacent sub-matrix, only the one-way edge from the node of the previous layer to the node of the next layer is reserved (reflecting the one-way evolution logic of the disease course), and the element weight is calculated by combining the time decay coefficient (the influence of the previous layer event on the next layer decreases as the time interval increases, such as the influence of acute phase treatment on follow-up period outcome decays by 30%) and the clinical stage transfer probability (such as the probability of heart function improvement in the recovery period after successful PCI in the acute phase). For example, the inter-layer edge weight between the nodes of "in-hospital acute phase → PCI treatment" and "recovery period → heart function stable" needs to refer to the clinical statistical rate of heart function improvement after PCI and the time decay effect. The whole adjacent matrix is designed by block and direction weight, which not only distinguishes the association difference of different disease course stages, but also connects the dependence transmission across stages, providing a structured input for the subsequent multi-modal spatio-temporal graph neural network model (MST-GNN) to learn dynamic hidden dependence.

[0063] It can be understood that the multi-modal features include structured features, image features, gene features and physiological features, and the timestamp feature can be understood as the time length of the node event occurrence time from the beginning of the disease course.

[0064] Further, the core layer of the multi-modal spatio-temporal graph neural network model includes time feature learning, spatial correlation learning and cross-modal fusion. In the time feature learning, 1D convolution (Conv1D) is used to extract the time sequence dependence of node events (such as the time interval feature of hypoglycemic drug use and blood glucose control), and a time sequence feature vector is output. In the spatial correlation learning, the spatial correlation of nodes in the same layer / cross layer is learned based on the graph attention network, and multi-modal attention weight is introduced, and a spatial feature vector is output. In the cross-modal fusion, the time sequence feature vector and the spatial feature vector are combined through a gating fusion unit, which can be represented as: fusion feature = σ (time sequence weight × time sequence feature + space weight × space feature), where σ is a sigmoid function, and the weight is adaptively learned through model training.

[0065] Further, the multi-modal correlation score reflects the dependence strength of the node and other nodes. In the embodiment of the present application, the top 20% of the core dependence path is first screened out based on the multi-modal correlation score, and the efficiency is improved by preferentially verifying the causality, avoiding meaningless low-intensity path verification.

[0066] Step S04, according to clinical knowledge, the nodes of the disease course hierarchical network are constrained, and the target path in the core dependent path is screened out.

[0067] Specifically, according to clinical knowledge, such as medical guidelines, expert experience, etc., the causal relationship of the node pair in the disease course hierarchical network is marked. When the causal relationship is clear causality, positive constraint is performed, that is, the initial value of the corresponding edge weight is increased. Illustratively, high blood pressure ACEI class drugs blood pressure drop belongs to clear causality. When the causal relationship is clear irrelevant, reverse constraint is performed, that is, the upper limit of the corresponding edge weight is set to filter the pseudo-association. Illustratively, cold antibiotic blood sugar rise belongs to clear irrelevant, and the invalid workload of subsequent causal verification is further reduced by filtering the pseudo-association. In some other embodiments of the present application, the constraint rule embedding also includes logical constraint, that is, the reverse time sequence dependence such as "treatment diagnosis" and "outcome treatment" is prohibited. Illustratively, blood pressure drop ACEI class drugs do not conform to the clinical logic and are automatically excluded.

[0068] It can be understood that the above steps further mark the path with high association score but marked as clear irrelevant as a suspected pseudo-association, reduce the invalid workload of subsequent causal verification, and screen the target path. At the same time, by adjusting the edge weight, the accuracy of model learning is ensured.

[0069] Step S05, counterfactual simulation verification is performed on the target path, and the final path that passes the verification is output.

[0070] Specifically, a clinical contraindication library is obtained, which integrates guideline contraindications, drug instruction contraindications, expert consensus contraindications, etc. The virtual intervention scheme for the target path is checked whether it conforms to the clinical operation specification. It can be understood that the virtual intervention scheme refers to the virtual adjustment of the parameters (such as dose, timing) of the screened target path, illustratively, diagnosis A treatment B outcome C. The change rate of outcome C is calculated, illustratively, treatment dose increase 20% outcome C improvement rate increase 15%. When checking whether the virtual intervention scheme for the target path conforms to the clinical operation specification, if the virtual intervention does not trigger any item in the contraindication library, it is qualified; if the contraindication is triggered, it is directly determined as unqualified, and the dependent path is excluded. Illustratively, for the path "heart infarction 2-24h PCI risk of reinfarction reduction", the virtual intervention "PCI is implemented for patients without PCI contraindications (such as severe bleeding tendency)", which does not violate the contraindication, and this dimension is qualified.

[0071] If the virtual intervention scheme for the target path is verified to comply with the clinical operation specification, the effect of the intervention on the outcome is verified by calculating the intervention effect. Specifically, for non-continuous variables, the absolute change value and the relative change percentage of the outcome after the intervention are calculated, i.e., the absolute effect and the relative effect. For example, after PCI intervention, the incidence of reinfarction is reduced from 15% to 8%, the absolute effect is 7%, and the relative change percentage is 46.7%. The relative effect is required to be greater than 20% in general, and the absolute change value and the relative change percentage are required to be greater than the corresponding threshold value. Meanwhile, for continuous variables, it is verified whether the intervention intensity is positively correlated with the improvement of the outcome or is stepwise correlated. For example, "the risk of reinfarction is reduced by 5% when the PCI timing is advanced from 24h to 12h, and the risk is further reduced by 4% when the timing is advanced to 6h", which complies with the clinical expectation that "the more optimal the timing, the stronger the effect", i.e., qualified. If "the dose is increased but the effect does not change" (e.g., the dose of anticoagulant is doubled, but the risk of bleeding does not increase), it is determined that the effect is unstable and unqualified.

[0072] If the absolute change value and the relative change percentage are greater than the corresponding threshold value, and the intervention intensity is positively correlated with the improvement of the outcome, the stability of the intervention effect under different conditions is verified by sensitivity analysis. Specifically, it includes verifying whether the effect fluctuation range of the intervention effect recalculated after adjusting the distribution range of the core confounding factors (such as age, underlying disease, and combined medication) (usually ±20%, such as "the mean age is adjusted from 65 to 52-78") is less than the preset fluctuation range, and verifying whether the Bootstrap sampling verification is up to standard. It should be noted that if the effect fluctuation range is less than 10% (e.g., the original absolute effect is 7%, and the fluctuation range is 6.5%-7.5% after adjustment), it indicates that the effect is not significantly disturbed by the confounding factors, and it is qualified. If the fluctuation is greater than 15% (e.g., it is reduced to 4%-10%), it is determined that the effect is unstable and unqualified. In addition, 1000 Bootstrap samplings (each sampling accounts for 80% of the original sample size) are performed on the simulation data, and the intervention effect of each sampling is calculated. If the 95% confidence interval does not contain "0" (e.g., the 95% CI is [5.2%, 8.8%], and "0" is not contained), and the interval width is less than 5% (e.g., the interval width is 3.6% in the above example), it indicates that the effect has statistical reliability, and it is qualified.

[0073] If the effect fluctuation range of the intervention effect recalculated after adjusting the confounding factors is less than the preset fluctuation range, and the Bootstrap sampling verification is up to standard, the causal logic of the intervention and the outcome is verified according to the medical pathological mechanism and the pharmacological principle to see whether it is contradictory.

[0074] If the verification intervention is not contradictory to the causal logic of the outcome, the corresponding path is output, for example, for the "diabetes → SGLT2 inhibitor → stable kidney function" path, the virtual intervention reduces the rate of kidney function deterioration, which can be explained by the mechanism of "SGLT2 inhibitor reduces kidney glucose reabsorption → reduces kidney load", the logic is consistent, and this dimension is qualified.

[0075] It can be found that only the verification of the four dimensions of clinical contraindication, effect quantization, stable effect, and logical consistency is qualified, and the corresponding "diagnosis-treatment-outcome" dependent path can be included in the clinical decision reference library; if any dimension is not satisfied, the path needs to be optimized again (such as adjusting the intervention parameters, supplementing multi-modal features) and verified again.

[0076] To sum up, the big data-based disease group data structure analysis method in the above embodiments of the application, the method obtains disease group data of a user, performs structured feature processing and unstructured feature processing on the disease group data, and splices the results of the structured feature processing and the unstructured feature processing through a cross-modal attention mechanism to form multi-modal features; according to clinical guideline stage division and data-driven risk inflection points, disease course stratification is determined, and a disease course stratification network is constructed, each layer of the disease course stratification network includes nodes exclusive to the corresponding stage, and the nodes include diagnosis nodes, treatment nodes, and outcome nodes; an adjacency matrix of the disease course stratification network, multi-modal features, and timestamp features are input into a multi-modal spatio-temporal graph neural network model, and multi-modal association scores of each node and core dependent paths of each layer of the network are output; according to clinical knowledge, the nodes of the disease course stratification network are constrained, and target paths in the core dependent paths are screened out; the target paths are subjected to counterfactual simulation verification, and a final path that passes the verification is output. Specifically, the application breaks through the limitation of a single-modal static network, fuses multi-dimensional medical data, constructs a causal reasoning system with dynamic stratification of disease courses, multi-modal feature fusion, and clinical knowledge constraints, realizes the leap from data association to clinically interpretable causal dependence, and directly supports personalized diagnosis and treatment decisions.

[0077] Embodiment Two

[0078] Please refer to Figure 2 , Figure 2 is a structural diagram of a big data-based disease group data structure analysis system provided by the embodiment two of the application. The big data-based disease group data structure analysis system 200 is used to realize the above embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that realizes a predetermined function. Although the devices described in the following embodiments are preferably realized in software, hardware, or a combination of software and hardware is also possible and conceived.

[0079] Specifically, the disease group data structure analysis system 200 based on big data comprises a feature processing module 21, a construction module 22, an input module 23, a screening module 24 and a verification module 25, wherein:

[0080] The feature processing module 21 is configured to acquire disease group data of a user, perform structured feature processing and unstructured feature processing on the disease group data, and splice results of the structured feature processing and the unstructured feature processing through a cross-modal attention mechanism to form multi-modal features, wherein in the structured feature processing, ICD-10 diagnosis codes, ATC medication codes and outcome indicators are converted into 64-dimensional feature vectors through coding embedding and numerical normalization processing;

[0081] In the unstructured feature processing, a medical pre-training model is used to extract lesion features to generate 256-dimensional image feature vectors;

[0082] In the unstructured feature processing, core gene sites related to diseases are screened through gene pathway analysis and converted into 128-dimensional gene feature vectors;

[0083] In the unstructured feature processing, a time convolution network is used to extract rhythm features to generate 64-dimensional physiological feature vectors;

[0084] The construction module 22 is configured to determine disease course stratification and construct a disease course stratification network according to clinical guideline stage division and data-driven risk inflection points, wherein each layer of the disease course stratification network comprises nodes exclusive to corresponding stages, and the nodes include diagnosis nodes, treatment nodes and outcome nodes;

[0085] The input module 23 is configured to input an adjacency matrix of the disease course stratification network, the multi-modal features and timestamp features into a multi-modal spatio-temporal graph neural network model to output multi-modal correlation scores of each node and core dependency paths of each layer of the network, wherein a core layer of the multi-modal spatio-temporal graph neural network model comprises time feature learning, spatial correlation learning and cross-modal fusion, in the time feature learning, 1D convolution is used to extract time sequence dependencies of node events to output time sequence feature vectors, in the spatial correlation learning, graph attention networks are used to learn spatial correlations of nodes in the same layer / cross layers and multi-modal attention weights are introduced to output spatial feature vectors, and in the cross-modal fusion, the time sequence feature vectors and the spatial feature vectors are combined through a gating fusion unit;

[0086] The screening module 24 is configured to constrain the nodes of the disease course hierarchical network according to clinical knowledge, and screen out the target path in the core dependent path, wherein the causal relationship between the nodes in the disease course hierarchical network is marked first, when the causal relationship is clear causality, forward constraint is performed, that is, the initial value of the corresponding edge weight is increased, and when the causal relationship is clear irrelevant, reverse constraint is performed, that is, the upper limit of the corresponding edge weight is set to filter pseudo-association.

[0087] The verification module 25 is configured to perform counterfactual simulation verification on the target path, and output a final path that passes the verification.

[0088] Further, in some optional embodiments of the present application, the construction module 22 comprises:

[0089] The division unit is configured to perform stage division according to a clinical guideline, and obtain a preliminary hierarchical result.

[0090] The identification unit is configured to perform LSTM-based time sequence inflection point identification on the historical data, correct the boundary of the preliminary hierarchical result, and determine the disease course hierarchy.

[0091] Further, in some optional embodiments of the present application, the verification module 25 comprises:

[0092] The checking unit is configured to obtain a clinical contraindication library, and check whether the virtual intervention scheme for the target path conforms to the clinical operation specification.

[0093] The first verification unit is configured to, if the virtual intervention scheme for the target path conforms to the clinical operation specification, verify the influence of the intervention on the outcome by calculating the intervention effect, specifically, for the case that the intervention is a non-continuous variable, calculating the absolute change value and the relative change percentage of the outcome after the intervention, and judging whether the absolute change value and the relative change percentage are greater than the corresponding threshold value, and for the case that the intervention is a continuous variable, verifying whether the intervention intensity and the outcome improvement are positively correlated / stepwise correlated.

[0094] The second verification unit is configured to, if the absolute change value and the relative change percentage are greater than the corresponding threshold value, and the intervention intensity and the outcome improvement are positively correlated / stepwise correlated, verify the stability of the intervention effect under different conditions by sensitivity analysis, specifically, including verifying whether the effect fluctuation amplitude of the intervention effect recalculated after adjustment of confounding factors is less than a preset fluctuation amplitude, and verifying whether Bootstrap sampling verification is up to standard.

[0095] The third verification unit is configured to verify whether the intervention and the outcome are contradictory in causality logic according to medical pathological mechanism and pharmacological principle if the effect fluctuation amplitude of the intervention effect re-calculated after the mixed factor adjustment is less than the preset fluctuation amplitude and the Bootstrap sampling verification meets the standard.

[0096] The output unit is configured to output the corresponding path if the intervention and the outcome are not contradictory in causality logic.

[0097] Embodiment three

[0098] Another aspect of the present application also provides an electronic device, referring to Figure 3 , which is an electronic device in the embodiment three of the present application, comprising a memory 20, a processor 10, and a computer program 30 stored in the memory and capable of running on the processor, wherein the processor 10 implements the big data-based disease group data structure analysis method as described above when executing the computer program 30.

[0099] In some embodiments, the processor 10 can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip, which is used to run program codes or process data stored in the memory 20, such as executing access restriction programs.

[0100] The memory 20 comprises at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory, etc.), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 20 can be an internal storage unit of the electronic device, such as a hard disk of the electronic device. In other embodiments, the memory 20 can also be an external storage device of the electronic device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 20 can include both the internal storage unit and the external storage device of the electronic device. The memory 20 can be used to store application software and various data of the electronic device, and also can be used to temporarily store data that has been output or will be output.

[0101] It should be noted that Figure 3 The structure shown does not constitute a limitation on the electronic device, and in other embodiments, the electronic device can include fewer or more components than shown, or combine certain components, or different component arrangements.

[0102] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the big data-based disease group data structure analysis method.

[0103] Those skilled in the art can understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a list of executable instructions for implementing the logic function, and can be specifically implemented in any computer readable medium for use by an instruction execution system, device or apparatus, such as a computer-based system, a system including a processor, or other system that can fetch and execute instructions from the instruction execution system, device or apparatus, or in conjunction with these instructions execution system, device or apparatus. For the present specification, the "computer readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by an instruction execution system, device or apparatus, or in conjunction with these instruction execution system, device or apparatus.

[0104] More specific examples (a non-exhaustive list) of the computer readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer readable medium can even be paper or other suitable medium on which the program is printed, because the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by electronic conversion, interpretation or processing, if necessary, in other suitable manner, and then stored in a computer memory.

[0105] It should be understood that parts of the present application can be realized in hardware, software, firmware or a combination thereof. In the above-described embodiments, a plurality of steps or methods can be realized by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if realized by hardware, and as in another embodiment, any one or a combination of the following technologies known in the art can be used: discrete logic circuit with logic gate circuit for implementing logic function on data signal, application specific integrated circuit with suitable combination logic gate circuit, programmable gate array (PGA), field programmable gate array (FPGA), etc.

[0106] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in one or more embodiments or examples.

[0107] The above embodiments only express several implementation manners of the present application, which are described in a more specific and detailed manner, but cannot be understood as the limitation of the patent scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for analyzing disease group data structure based on big data, characterized in that, The method comprises: acquiring disease group data of a user, performing structured feature processing and unstructured feature processing on the disease group data, and splicing results of the structured feature processing and the unstructured feature processing through a cross-modal attention mechanism to form multi-modal features; determining disease course stratification according to clinical guideline stage division and data-driven risk inflection points, and constructing a disease course stratification network, each layer of the disease course stratification network comprising nodes exclusive to a corresponding stage, the nodes including diagnosis nodes, treatment nodes, and outcome nodes; inputting an adjacency matrix of the disease course stratification network, the multi-modal features, and timestamp features into a multi-modal spatio-temporal graph neural network model to output multi-modal correlation scores of each node and core dependency paths of each layer of the network; constraining the nodes of the disease course stratification network according to clinical knowledge, and screening target paths in the core dependency paths; performing counterfactual simulation verification on the target paths to output a final path that passes verification; in the step of constraining the nodes of the disease course stratification network according to clinical knowledge and screening target paths in the core dependency paths, first, mark the causal relationship of a node pair in the disease course stratification network, when the causal relationship is clear causality, perform positive constraint, that is, increase the initial value of the corresponding edge weight, and when the causal relationship is clear irrelevance, perform reverse constraint, that is, set an upper limit of the corresponding edge weight to filter pseudo-association; the step of performing counterfactual simulation verification on the target paths to output a final path that passes verification comprises: acquiring a clinical contraindication library to check whether a virtual intervention scheme for the target path conforms to clinical operation specifications; if the virtual intervention scheme for the target path conforms to the clinical operation specifications, verify the influence of the intervention on the outcome by calculating the intervention effect, specifically, for non-continuous variables, calculate the absolute change value and the relative change percentage of the outcome after the intervention, and determine whether the absolute change value and the relative change percentage are greater than the corresponding threshold value, and at the same time, for continuous variables, verify whether the intervention intensity and the outcome improvement are positively correlated / stepwise correlated; if the absolute change value and the relative change percentage are greater than the corresponding threshold value, and the intervention intensity and the outcome improvement are positively correlated / stepwise correlated, verify the stability of the intervention effect under different conditions through sensitivity analysis, specifically, including verifying whether the effect fluctuation amplitude of the intervention effect recalculated after adjustment of confounding factors is less than a preset fluctuation amplitude, and verifying whether Bootstrap sampling verification is up to standard; if the effect fluctuation amplitude of the intervention effect recalculated after adjustment of confounding factors is less than the preset fluctuation amplitude, and the Bootstrap sampling verification is up to standard, verify whether the causal logic of the intervention and the outcome is contradictory according to medical pathological mechanisms and pharmacological principles; if the causal logic of the intervention and the outcome is not contradictory, output the corresponding path.

2. The big data based patient group data structure analysis method of claim 1, wherein, In the step of obtaining the user's disease group data, structuring feature processing and unstructured feature processing of the disease group data, and splicing the results of the structured feature processing and the unstructured feature processing through a cross-modal attention mechanism to form multi-modal features, in the process of structured feature processing, the ICD-10 diagnosis code, ATC medication code, and outcome indicators are converted into 64-dimensional feature vectors through coding embedding and numerical normalization processing. In the process of unstructured feature processing, a medical pre-training model is used to extract lesion features to generate a 256-dimensional image feature vector. In the process of unstructured feature processing, core gene sites related to the disease are filtered through gene pathway analysis and converted into 128-dimensional gene feature vectors. In the process of unstructured feature processing, a time convolution network is used to extract rhythm features to generate a 64-dimensional physiological feature vector. 3.The big data-based patient group data structure analysis method of claim 2, wherein, The step of determining the disease course stratification according to the clinical guideline stage division and the data-driven risk inflection point includes: According to the clinical guidelines, the stage division is performed to obtain a preliminary stratification result; The historical data is subjected to LSTM-based time series inflection point identification to correct the boundaries of the preliminary stratification result and determine the disease course stratification.

4. The big data based patient group data structure analysis method of claim 3, wherein, Different hierarchical nodes are connected through stage transition edges, wherein the edge weight of the stage transition edge is the contribution degree of the corresponding treatment to the stage transition, which is calculated through Cox regression.

5. The big data based patient group data structure analysis method of claim 4, wherein, The core layer of the multi-modal spatio-temporal graph neural network model includes time feature learning, spatial correlation learning, and cross-modal fusion. In the time feature learning, 1D convolution is used to extract the time sequence dependence of node events to output a time sequence feature vector. In the spatial correlation learning, the spatial correlation of nodes in the same layer / cross layer is learned based on a graph attention network, and a multi-modal attention weight is introduced to output a spatial feature vector. In the cross-modal fusion, the time sequence feature vector and the spatial feature vector are combined through a gating fusion unit. 6.A big data-based disease group data structure analysis system, characterized by comprising: The system for implementing the big data-based disease group data structure analysis method according to any one of claims 1-5 includes: A feature processing module for obtaining the user's disease group data, structuring feature processing and unstructured feature processing of the disease group data, and splicing the results of the structured feature processing and the unstructured feature processing through a cross-modal attention mechanism to form multi-modal features; A construction module for determining the disease course stratification according to the clinical guideline stage division and the data-driven risk inflection point, and constructing a disease course stratification network, wherein each network of the disease course stratification network contains nodes exclusive to the corresponding stage, including diagnosis nodes, treatment nodes, and outcome nodes; An input module for inputting the adjacency matrix of the disease course stratification network, the multi-modal features, and the timestamp features into a multi-modal spatio-temporal graph neural network model to output multi-modal correlation scores of each node and core dependency paths of each network; A screening module for constraining the nodes of the disease course stratification network according to clinical knowledge and screening target paths in the core dependency paths; A verification module for performing counterfactual simulation verification on the target paths to output a final path that passes the verification.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by the processor, implements the big data-based disease group data structure analysis method according to any one of claims 1-5.

8. An electronic device, comprising: A computer program product comprising a memory, a processor, and a computer program stored on the memory and loadable on the processor, the processor implementing the big data-based disease group data structure analysis method according to any one of claims 1-5 when executing the program.

Images